Mistral Launches Agents API Framework for Developing AI Agents

The French artificial intelligence company Mistral has introduced a new framework known as Agents API, designed to enable organizations to develop their own AI agents for task and process automation. This innovative framework integrates both Mistral's proprietary language models and external data sources, providing a robust control mechanism for specialized AI agents along with a persistent memory feature that maintains context across multiple agents.

One of the standout features of the Agents API is its array of built-in connectors that facilitate various functionalities, such as executing code or accessing documents. To interface with external applications, the framework employs the Model Context Protocol (MCP) created by Anthropic.

Currently, the framework includes four essential connectors that can be accessed freely by AI agents. Users can also utilize these connectors directly through a chatbot interface. Among these connectors is a code interpreter capable of executing Python code in a secure sandbox environment, where it can also check for syntax errors. This feature is useful for performing mathematical calculations and conducting data analysis and visualization. Additionally, through integration with GitHub and the recently updated Devstral language model, users can create coding assistants.

Moreover, the Agents API includes a connector for image generation, leveraging the Flux 1.1 Pro Ultra language model from Black Forest Lab. According to Mistral, this functionality allows users to create visual aids for training materials or marketing graphics. The document library connector enables AI agents to access documents stored in the Mistral cloud, facilitating Retrieval Augmented Generation (RAG) by utilizing document content as contextual information for generating outputs.

Another significant enhancement is the built-in web search capability, which supplements the training data of the language models with up-to-date information. AI agents can autonomously search the web based on user inquiries, and users can also provide specific URLs to retrieve information from targeted websites. In performance benchmarks, the inclusion of web search dramatically improved the quality of responses from Mistral's Medium and Large language models, with success rates tripling. The Mistral Medium model achieved a score of 75 percent, while the larger model reached 82 percent.

Developers aiming to create custom workflows involving multiple AI agents can start by defining all the necessary agents, each with access to different language models and applications. There is no limit to the number of agents that can be utilized within the Agents API. Developers can then decide whether a specific AI agent should produce an output or delegate a request to another agent, allowing for the chaining of multiple agents. For instance, a financial analysis agent can delegate the research of stock market data to a web search agent and pass the calculations to another specialized agent.

For those interested in exploring the Agents API, detailed documentation is available, along with a selection of pre-built example agents provided by Mistral on GitHub. Initially, the framework supports the latest versions of Mistral's Medium and Large language models, with plans to incorporate additional models in the future. Recently, JetBrains also released the Koog framework, an open-source tool designed for modular AI agent development.